journal title a data-driven approach to agent based ... · simulation (abms) as a novel approach to...

A Data-driven Approach to Agent BasedExploration of Customer Behaviour

Journal TitleXX(X):1–13c©The Author(s) 2016

Reprints and permission:sagepub.co.uk/journalsPermissions.navDOI: 10.1177/ToBeAssignedwww.sagepub.com/

David Bell1 and Chidozie Mgbemena2

AbstractCustomer retention is a critical concern for most mobile network operators because of the increasing competition inthe mobile services sector. This concern has driven companies to exploit data as an avenue to better understandcustomer needs. Data mining techniques such as clustering and classification have been adopted to understandcustomer retention in the mobile services industry. However, the effectiveness of these techniques is debatable due tothe increasing complexity of the mobile market itself. This study proposes an application of Agent-Based Modeling andSimulation (ABMS) as a novel approach to understanding customer retention. A dataset provided by a mobile networkoperator is utilised to automate decision trees and agent based models. The most popular churn modeling techniqueswere adopted in order to automate the development of models, from decision trees, and subsequently explore customerchurn scenarios. ABMS is used to understand the behavior of customers and detect possible reasons why customerschurned or stayed with their respective mobile network operators. Data analysis is able to identify that location andchoice of mobile devices were determinants for the decision to churn or stay with their mobile network operator - withword of mouth as an important factor. Importantly, agent based simulation is able to explore further the determinants inthe wider marketplace.

KeywordsAgent Based Modelling; CADET Approach; Decision Trees

Introduction

Simulation models that describe agents have become awidely used tool for understanding phenomena. Thesemodels have been used across industries and often provideinsights into complex problems. A popular simulation modelis the Agent-based model (ABM). The agent-based modelallows researchers and practitioners to study how systemlevel properties emerge from the adaptive behaviour ofindividuals and on the other hand how systems affectindividuals (1).

Agent-based models consist of a number of entities withindividual rules of behaviours. Entities in such modelsinteract with one another and with their surroundingenvironment. Such interaction may influence the behaviourof the agents. Harnessing this information and understandingthe influence of agents interaction with other agents andagent interaction with the environment can provide someuseful insights to business problems, in this case customerchurn.

There are various ways of describing agent based models.Adopting the design science paradigm, this paper presentsa novel approach to agent based modelling. The CustomerAgent Decision Tree (CADET) approach is a data-drivenapproach that provides key drivers that collectively uncoverthe decision of individual agents in an environment. TheCADET approach is not industry specific. Thus, it canbe applied in different sectors. The CADET approach isvalidated by applying it to investigate customer retention inthe mobile services industry (MSI). The next section presentsa background on customer retention in the MSI

Background on Customer RetentionOver the last decade, the number of mobile phone users haveincreased reaching an overwhelming number of 7 billion (2).In developed countries, telecommunication companies havemobile penetration rates above 100% with no new customers(3). Therefore, customer retention receives a growing amountof attention from telecommunication companies. The highpenetration rate of above 100% in the mobile servicessector has motivated substantial research into customerretention. It has been shown in the literature that customerretention is profitable to a company because: (1) acquiringnew customers cost five times more than retaining existingcustomers (4; 5). (2) Existing customers generate higherprofits, become less costly to serve, and may provide newreferrals through providing positive word-of-mouth whiledissatisfied customers might spread negative word of mouth(6; 7; 8). (3) Loosing customers may lead to opportunitycosts because of reduced sales (9; 10). However, a smallimprovement in customer retention can lead to a significantincrease in profit (11). A number of studies in the area ofcustomer retention have revealed that customer satisfactionis a strong predictor of customer retention (12). This isas a result of the relationship between both conceptualelements (13). Factors that drive customer satisfaction

Corresponding author:David Bell, Brunel University, Kingston Lane, Uxbridge, London, UB83PH, UK.Email: [email protected]

Prepared using sagej.cls [Version: 2015/06/09 v1.01]

2 Journal Title XX(X)

(such as service quality) can also drive customer retention.(14). Furthermore, customer satisfaction is often seen as amotivator for customer retention (15). While it seems thatsatisfied customers will remain with their mobile serviceprovider, this is not always the case because satisfiedcustomers can defect while dissatisfied customers can beretained (15).

Customer churn is a term that is widely used in the areaof customer retention to describe customers who switch to adifferent mobile service provider or leave the market entirely.

There are two basic approaches that can be used to addresscustomer churn namely, untargeted and targeted approaches(16). Untargeted approaches rely on outstanding productand mass advertising to increase brand loyalty and retaincustomers while targeted approaches rely on identifyingcustomers who are likely to churn and then providing themwith either a direct incentive or a customised plan for themto stay (16). Various types of information can also be usedto predict customer churn, such as information on socio-demographic data (e.g. sex, age, or post code) and callbehaviour statistics (e.g. the number of international calls,billing information, or the number of calls to the customerhelpdesk).

The main factors that influence customer churn in themobile services market are (1) customer satisfaction, (2)switching costs, (3) relationship quality and (4) price (17;18; 19). Price is the most important factor for customerchurn, followed by customer service, service quality andcoverage quality (20). However, social influence is anotherkey driver to customer churn in the mobile services industry(21; 22). This paper proposes an approach to ABM in orderto explore social influence on customer churn behaviour.The next section presents approaches to modelling customerbehaviour with respect to customer retention.

Customer Modelling BehaviourCustomers often interact with other customers aboutproducts and services they purchase. In addition to theadvertisement campaigns carried out by MNOs, interactionamong customers also influences customer purchase andrepurchase decision. There are various theories in socialscience and marketing concerned with understanding andmodelling customer behaviour (23). Customer behaviourmay change as a result of an act of a consumer changingpreference on a product or service (24). A number ofapproaches have been applied to understand the conceptbehind consumers changing preferences (24).

Customer retention can be achieved if a company isable to understand patterns in which customers behave andthe likely triggers for such behaviour (25). Understandingcustomers, managing interactions and relationships withthem is a vital part of customer relationship management(CRM). A company with a good CRM should be ableto predict possible changes in customer behaviour (25).Predicting customer behaviour can be achieved throughcustomer behaviour modelling (26), by applying tools andtechniques to gain a better insight on customer behaviouralpatterns and in turn predict future behaviour (27). Neslin et.al., (28) characterised CRM models as either analytical orbehavioural models. Analytical models involve large datasets

that are stored in data warehouses. These datasets requiremodels that can easily scale the dataset and provide resultsto increase company revenue (28). Behavioural modelsmake use of surveys to analyse cognitive responses toservices provided (28). Furness (29) classified customerbehaviour modelling into (1) descriptive modelling, (2)predictive modelling and (3) a combination of descriptiveand predictive modelling. Descriptive modelling describesmodels that attempt to answer why questions. An exampleof descriptive modelling is clustering. When a customerclustering exercise is conducted, customers belong to acertain cluster because they collectively possess similarattributes or behaviours. Predictive modelling describesmodels that answer the who questions. For example whowill buy a product or service? In the context of customerchurn, predictive models can give insight on who islikely to churn (30). Predictive models typically work bypredicting future customer behaviour based on their pastbehaviour (30). Finally, the combination of descriptive andpredictive modelling addresses problems by integrating bothdescriptive and predictive models to provide a more conciseanswer on who and why questions at the same time (3).Descriptive and predictive models are popularly carried outusing data mining approaches. The next section presents anoverview of applying data mining approaches to investigatecustomer churn.

Applying Data Mining Techniques for ModelingChurnData mining techniques are popularly used to build churnprediction models (31; 32). Data mining simply meansextracting hidden knowledge from data, and it is a populartechnique for understanding customer behaviour from rawdata. Data mining methods are widely used in the literaturefor analysing and investigating customer churn for two mainreasons: they have better prediction results (33) and they aremore suitable for analysing large data sets (34).

Numerous studies have applied various data miningtechniques to study customer churn. These studies include(35; 36; 37). The most popular data mining techniques forpredicting churn include decision trees, logistic regressionand support vector machines (SVM) (35). Furthermore,numerous industries have attempted to apply data miningtechniques to study customer churn. These industries includebanking (38; 39), insurance (40), retail (41; 42), andeconomics (28). Most studies that addressed customer churnusing data mining techniques do not account for the socialeffect of customer retention (43). As a result, this areahas been neglected and it is considered a major driver forcustomer retention (21; 22). The next section presents anoverview of the social network analysis.

Social Network AnalysisSocial network analysis (SNA) is an evolving scientificresearch area. SNA has developed to be a primary techniquefor describing the social structure and interaction betweennetwork representation (44). A social network is an inter-connection between nodes using various links (24). For thepurpose of this study, nodes represent customers and thelinks represent the relationships between customers.

Prepared using sagej.cls

Mgbemena and Bell 3

SNA typically focuses on static networks. Static networksare the mapping of relationships between discrete entities(45). These networks do not change their structure over time.Recently, researchers have focused on dynamic networks thatare capable of representing the continuous transmission ofinformation and influence (46). Dynamic networks are a vitalaspect of ABMS. Using dynamic network analysis involvesunderstanding the agent rules that govern network structureand growth, and how networks and their relationships conveyinformation (47). SNA is an approach to anticipating andmodelling society as different sets of people or groups linkedto one another (24). SNA is a method of enquiry thatfocuses on the relationships between subjects. This approachseeks to understand subjects by collecting information fromdifferent sources, analysing the information and visualisingthe results. SNA has proved to be useful in explainingsome important phenomena such as investigating the spreadof disease, understanding the internet and explaining thesmall word effect to the spreading of information (48).Sociologists and market researchers believe that the life ofan individual depends on how that individual engages withthe web of social connections (45). The social networks termis loosely used to refer to social and professional networkingsites including but not restricted to Facebook, LinkedInand Twitter. Each networking site represents an onlinecommunity. An online community is an example of a socialstructure. The people (also known as the nodes) who signup on these online communities alongside the relationshipbetween them represent social networks. A social networkstructure consists of nodes and the interrelationship betweenthe nodes. Due to the growing interest of social networkanalysis, researchers have investigated the principles of thenetwork approach. These principles include (24);

1. Actors and their actions should be viewed asautonomous and independent units rather than asinterdependent units.

2. The links between actors should be viewed as channelsfor transfer of material and non-material resources.

3. Social network models focusing on individuals viewthe structural environment as a network imposingcertain constraints on individual actions

4. Social network models conceptualise structure (social,economic, political, and so on) as long-lasting patternsof relations among actors.

Milgrim (48) is an empirical research study of socialstructure and it introduces ’six degrees of separation’phenomenon while addressing the ’small world problem’.In this study, some participants were chosen at random andasked to deliver a letter to a target person using only a chainof friends and acquaintances. There was a starting person anda target person in different states in America. The startingperson was advised to send the letter across to his friendor acquaintance who is likely to know the target person.The procedure started in Nebraska and the destination of theletter was Boston, Massachusetts. The process will continueuntil the message gets to the target person. When the lettersarrived in Boston, Milgram discovered that it took an averagenumber of six steps for the letter to get to the destination.Hence, Milgram labelled the phenomena six degrees ofseparation. This study concludes by describing how people

of a population are connected. Although (48) study wasnot subject to an evaluation process, his concept of thesmall world is widely adopted in social networks researchto provide an explanation about how information spreads inthe real world. The small-world network has become one ofthe most widely-used social network models (49). The nextsection provides a background on social impact of customerretention.

Social Impact on Customer RetentionMore recently, the influence of social networks have beenfound to be a key contributing factor to customer retention(50; 51). Social networks’ influence is typically carried outwith the use of word of mouth (WOM) (52). WOM is theinformal communication between private parties regardingevaluations of goods and services (53). 75% of defectedcustomers spread a negative WOM to one or more customers(54). The following section presents some relevant studies onthe impact of social influence on customer retention.

(21) developed a model that integrates social networkanalysis with traditional churn modelling concepts. Themodel was applied to a dataset of over half a millionsubscribers, provided by a large mobile network provider.The dataset contained customer call detail records (CDR). Tocompute social tie strength, the authors used three attributesnamely: 1) the number of calls placed between two users,2) the total duration of calls between two users and 3)neighbourhood overlap of the two connected users. Thisstudy found users that make phone calls to numbers on adifferent network are likely to churn in future to save costs.

Similarly, (22) conducted a study investigating the impactof social networks on customer retention. However, thelatter differs from the former in that it uses both networkedand non-networked (customer-related) information aboutmillions of users. The key finding of this study was that churnnot only had an impact on customers’ friends, it also had animpact on friends of friends.

Although the studies mentioned above have contributed tothe knowledge of customer retention, they do not capture thepossible factors as to why customers made their decision tochurn. ABMS is able to capture the dynamics of customerbehaviour (55). The next section presents a background onABMS.

Agent Based Modelling and SimulationOver the years, researchers and industry practitioners haveattempted to apply different techniques to understandcustomer behaviour in the market place. The ABMSapproach is a typical example of one of these techniques(56). ABMS provides an understanding of how systemswork under certain conditions. ABMS works by creatingscenarios that imitate real life conditions, for example,carrying out an ABMS exercise with customers who walkinto a retail store. ABMS can be used in this context to deriveinsights on customer behaviour in the retail store (57). Thisprocess provides an explanation of the relationship betweenelements in a complex system. ABMS is composed of twomain activities: modelling and simulation. Modelling is theprocess of representing real life events into a model whilesimulation is the process of executing the represented models



such that they imitate the proposed system. Agent basedmodels are composed of agents and a structure for agentbased interaction.

Agents can represent anything from a number of patientsin a hospital to consumers of a product or service.ABMs are often characterised by rules and these rulesdefine the behaviour of agents in the system (47). Thesebehaviours are often influenced by agent interactions withother agents in the system, making the outcome difficultto predict. In such cases, a balance may be difficult toreach, making the ability to study the underlying system andthe dynamics of the behaviour imperative. ABM is distinctfrom traditional modelling approaches where characteristicsare often aggregated and manipulated (49). Traditionalmodelling techniques for modelling maybe suitable for theirpurposes but they may not be able to provide adequate levelof details in regards to the independent behaviours of agents.Although the commonly used data mining techniques formodeling consumer markets are powerful with regard to theirpurposes, they are generally not able to provide sufficientlevels of detail with regards to the modelling interdependentbehaviors of consumers in a market place

In addition, ABM is able to sufficiently represent interde-pendent systems even on a large scale i.e. incorporating ahigh number of factors, with each factor’s level of detail andthe behavioral complexity of those factors (57). ABM is arelatively new approach for modelling complex systems thatare composed of interacting independent elements (47). TheABM technique can be applied to any aspect of a phenomena(47). ABM has been applied in various areas including eco-nomics (58), health-care (59), management science (47) andgeography (60). In business, ABM has been applied to helpdecision makers understand underlying market structuresand anticipate dynamics in the market place (47). ABM hasalso been utilised in artificial life research, to explore lifein order to uncover how it might be, rather than how itactually is (61). ABM is also used in consumer modellingto understand and to predict consumer modelling process(62). Consumers are represented as independent agents withindividual characteristics and an independent decision mak-ing process (63). Sellers are represented as agents whopresent their products with different characteristics into themarket (64). In the study of social behaviour and interactions,ABM starts with a set of assumptions derived from the realworld (deductive), and produces simulation-based data thatcan be analysed (inductive) (65). ABM must create a clearrepresentation of what happens in reality so that every agentperforms a task of an individual as if it is happening in socialreality (64).

ABM has a number of benefits such as the abilityto model individual decision making while incorporatingheterogeneity and interaction/feedback (66). In addition,ABM has the ability to incorporate social/ecologicalprocesses, structures, norms and institutional factors (67).These advantages make it possible to couple human andnatural systems in an ABM. This paper applies the designscience paradigm as an overarching methodology with theapplication of the CADET approach to capture the dynamicsin customer behaviour in a simulated environment. The nextsection provides a background on design science researchmethodology.

Design Science Research Methodology

Design science research is a multidisciplinary approach thatprimarily uses design as a research method or techniqueto solve a problem and learn from the process of solvingthat problem (68). Apart from its popular adoption ininformation systems, it is also widely used in disciplinessuch as education, engineering computer science, and health-care (69). March and Smith (70) describe DSR as a researchmethodology that allows research to produce relevant andimproved effectiveness by strategically combining researchoutput (product) and research processing (activities) fromboth natural and design science in a two-dimensionalframework. The design science output or artefacts includesconstructs, models, methods and instantiations while thenatural science activities include build, evaluate, theorise andjustify (70). The DSR output classification defined by (70)can help establish appropriate measures to build, evaluate,theorise and justify a DSR. The four research outputs aredescribed below:

Constructs: Constructs are a set of concepts that areused to describe problems within a domain and specifytheir solutions. Constructs also form the vocabulary of adiscipline.

Models: Are a set of statements which expressrelationships among constructs and represent real-worlddesign activities in a domain (70). Models can also be usedto suggest solutions to problems in a solution space.

Methods: Are a sequence of steps used to execute a task.These steps provide guidelines on how to solve problemswith the use of constructs and models. Furthermore, methodscan be described as a set of methodological tools that arecreated by design science and applied by natural science(70).

Instantiations: Are the utilisation of constructs, modelsand methods to showcase an artefact in a domain.They demonstrate the effectiveness of constructs, modelsand methods (70). Newell and Simon (71) describethe importance of instantiations in computer science byexplaining how they offer a better understanding of aproblem domain, and as a result, provide improved solutions.Instantiations provide working artefacts that can drivesignificant advancement improvement in both design andnatural sciences. A DSR methodology incorporates fivestages of a design cycle to address design research problem.These phases are designed to aid sustainable developmentduring the research and transfer knowledge from oneiteration to the next iteration until a desired result is achieved.The next section explains the DSR processes.

The Design Science Research ProcessThe DSR process follows a stepwise approach structured asfive phases.

Awareness of Problem : The DSR process begins byidentifying the problem under study. The identified problemmay arise from multiple sources such as the literature orcurrent problems in the industry. The research problem needsto be clearly defined and articulated. The output of this phaseis a formal or informal proposal for new research. In thisstudy, the main problem is identifying an effective approach


Mgbemena and Bell 5

Figure 1. Design Science Research Process

to modelling customer behaviour as a means to improvecustomer retention.

Suggestion: This phase is explored when a researchproposal has been presented. Possible solutions about theresearch problem are explored and evaluated, leading to theacquisition of further insights to the domain under study.The specifications of the appropriate solutions to the researchproblem are defined. The output of this phase is a conditionaldesign or representation of proposed solutions. In this study,the CADET approach to ABMS is suggested in this phase.

Development: This phase involves further developing andimplementing the DSR artefacts based on the suggestionsfrom the previous phases. The outputs of this phase are theartefacts, which are core elements of the DSR process. (70)described DSR artefacts into four categories: Constructs,models, methods and instantiations. In this phase, theCADET approach to ABMS is developed using a selectedmachine learning algorithm.

Evaluation: The developed artefacts are analysed andevaluated according to the criteria set in phase 1 (awarenessof problem phase). Deviations and expectations should benoted and explained in this phase. If the outcomes derivedfrom the development or evaluation phase do not meet theobjectives of the problem, the design cycle returns to the firstphase, along with the knowledge gained from the process ofthe first round of work. These phase may be iterated until theevaluation of the artefacts meets the solution requirements.The outputs of this phase are performance managementthat should improve the efficiency and effectiveness of theartefact. In this phase, the CADET approach to ABMS isevaluated by its application to a telco dataset on an ABMSplatform.

Conclusion: This is the last phase of the DSR cycle. Theresults of the research are written up and communicatedto a wider audience in forms of professional publicationsand scholarly publications (72). Kuechler and Vaishnavi (73)categorise the knowledge gained in this phase as either firmor loose ends. Firm knowledge are facts that have beenlearned and can be repeatably applied or behaviour thatcan be repeatably invoked, while loose ends are anomalousbehaviour that defies explanation and may well service asthe subject of further research (73). The CADET approachto ABMS can be categorised as firm knowledge because itcan be replicated on different datasets (both structured andunstructured) in various industries. The next section providesa summary of DSR evaluation.

Design Science Research EvaluationEvaluation is an integral part of the DSR process. Usually, itis concerned with answering the question ’How well does

the artefact work?’ (70). The evaluation process providesan avenue to validate the performance of an artefact andmeasure progress according to the defined metrics (70).Artefacts are constructed to carryout specific problems,thereby, demonstrating their effectiveness in solving theproblems. The process of developing an artefact may resultin deviations from expectations. In this case, these deviationsshould be properly explained (73). Knowledge gained fromthe evaluation phase of one iteration can be applied intofurther iterations. Evaluation plays an essential role in DSRas it is iterative in manner. Hence, it is important to developappropriate evaluation metrics to assess artefact performanceand to measure the efficiency and effectiveness of the artefactdeveloped (70). The criteria for evaluating the quality of anartefact depends on the artefact type (70).

Table 1 presents types of artefacts as described by (70)and their evaluation criteria. The main artefact derived in thisstudy is the CADET approach and it is evaluated by applyingit to a Telco dataset. The CADET approach was derived as ameans to address the crucial problem of customer retentionin the MSI. A number of studies in the literature have appliedABMS to address customer retention (74; 75), however,the CADET approach is different from other studies asit applies decision trees to model agent behaviour in anABMS environment. The next section presents an overviewof decision trees as a means to derive the CADET approach.

Decision TreesA decision tree is an analytical decision tool that usesa tree-like graph and their possible consequences toarrive at a decision. The tree-like structure represents therelationships between events. Decision tree analysis is amethod commonly used in data mining (76). The goal ofdecision trees is to create a model that predicts a targetvariable based on the input variables. Each leaf on a decisiontree shows a connection to the target variable.

Decision trees are sequential models that logicallycombine a series of simple tests; every individual testcompares a numeric attribute against a threshold value ora nominal attribute against a set of possible values (77).Decision trees consist of many nodes and branches indifferent stages and various conditions. (78). Decision treeanalysis are popularly used for customer churn analysisin the MSI (35). In the case of this study, instances areclassified into churn or no-churn. Decision tree models areusually represented in a top-down manner. The main aim ofa decision tree is to derive a tree which solves a particularbusiness problem and is easy to understand. To achieve suchresult, the decision tree undergoes two stages - tree buildingand tree pruning. Tree building is carried out using top-down



Name DescriptionArtefact Brief Description Evaluation CriteriaConstruct The conceptual vocabulary and symbols

describing a problem within a domainCompleteness, clarity, elegance, ease ofunderstanding and ease of use.

Model A set of propositions or statementsexpressing relationships between con-structs. Models represent situation as prob-lem and solution statements.

Precision with real-world phenomena,completeness, level of detail, robustness,and internal consistency.

Method A sequence of steps used to perform a task.A method can be tied to a particular model.A method may not be articulated explicitlybut represents tasks and results.

Operationality ( ability for the method tobe reused), efficiency, generality and easeof use.

Instantiations Application of constructs, models andmethods to provide working artefacts.

Efficiency and effectiveness of an artefacts.Also, the influence of the artefact on itsusers and on the environment at large.

Table 1. DR Artefact Evaluation Criteria (69)

strategy (also known as a divide and conquer strategy). Theprocess of tree building involves:

• Selecting the attribute for the root node.• Splitting instances into subsets.• Repeat recursively for each branch.• Stop if all instances have achieved the same class.

A root node is selected by comparing the number of bits(splitting based on information gain) for possible root nodesand choosing the node that has the most bits of information.After selecting the root node, the next step is to look atthe branches that emanate from the root node. The processof selecting the node with the most bits of information isrepeated. This process continues until all instances have thesame number of bits i.e. when there are no more classesto split on (accuracy is 100%). The tree pruning processinvolves eliminating error-prone branches. A pruned treecan improve a classifiers performance and can facilitatefurther analysis of the model for the purpose of knowledgeacquisition. The pruning process should never removepredictive parts of the classifier. For better understanding ofhow decision trees work, the example below (Figure 2):

If (Contract length >=24 months, Location= midlands,priceplan <30 and data <4gb), then Churn = Yes.

A number of algorithms can be used to build decisiontrees. These algorithms include CART (Classificationand Regression Trees), Chi-squared automatic interactiondetector (CHAID), Iterative Dichotomiser (ID3) and C4.5which is the successor of ID3. The CART algorithm isselected to build the CADET approach to ABM. The nextsection presents the CADET approach.

The CADET ApproachA number of approaches techniques can be used to carryoutABMS. Researchers have extended agent oriented method-ologies in two areas; object oriented (OO) methodologiesand knowledge engineering methodologies(79). The threecommon views in OO technologies for describing ABMs arestatic for describing the structure of objects; dynamic fordescribing object interaction and functional for describingthe data flow of the methods of the objects (79). The flow

Figure 2. Decision Tree Classification Process

Figure 3. CADET Utilisation Process

chat is another popular example for representing the processflow of an agent based model (1).We present the CADET approach, a novel approach for agentbased modelling using results derived from decision treeanalysis. The CADET approach is derived by applying theCART algorithm to a dataset and visualising the tree-likestructure derived from the analysis. The agent attributes arederived from the decision tree flow process. The CADETapproach can be applied to various domains and industries.


Mgbemena and Bell 7

Figure 4. An example of describing an ABM using the CADETapproach

To further explain the structure and functionality of theCADET approach, we present a scenario below;

A Mercedes Benz car dealership provides a number ofproducts and services. In order to understand customerpurchase behaviour, a decision tree analysis is conducted.The analysis is carried out to improve understanding oncustomer purchase behaviour. A decision tree approach ischosen to understand the trend and the pattern for the processof customer decision. The analysis shows that customers whohave certain attributes purchase Mercedes Benz cars fromthe Mercedes Benz car dealership. The variables used forthis analysis are age, height, salary, credit score, previousMercedes Benz purchase, number of services/repairs andsatisfaction level. The target variable is either to purchase aMercedes Benz car or to go purchase another car brand.

The decision tree analysis shows that customers that areover 30 years old, are often interested in purchasing aMercedes Benz while customers who are less than 30 yearsold often buy other car brands. The next variable on thedecision tree is height. Customers with height less than170cm often buy other brands while taller customers areoften interested in purchasing Mercedes Benz. The nextvariable on the decision tree is salary. If salary is greater than30,000, move to the next step, else consider purchasing otherbrands. The next step is credit score. If credit score is greaterthan 7, move to the next step otherwise, consider purchasingother brands. This process is followed until the end of thetree. Figure 3 displays a diagrammatic representation ofthe CADET framework and Figure 4 displays the decisiontree analysis process described above. The TEA-SIM tool isused to validate the CADET approach. Hence, it is brieflydescribed below.

TEA SIM ToolOver the years, companies have spent a large amount ofmoney on products and services that enable them manageand understand their customer behaviour more effectively.

Advertisement and word of mouth can be a powerfultools for customer retention. However, customers are oftenskeptical about advertisement and may turn to their family orfriends within their network to seek advice before decidingwhether to purchase or repurchase a product or service.Some companies have manipulated word of mouth byrunning schemes that offer customers benefits for expressingpositivity about their product or service. As positive word ofmouth is a great tool for marketing, negative word of mouthcan also be a damaging tool for businesses. Dissatisfiedcustomers of a product or service may share their experienceabout their dissatisfaction with members of their socialnetwork which can include family and friends.

The TEA-SIM tool is a data-driven agent based simulationplatform (80) built by incorporating a cognitive process forunderstanding how the members of a small-world network(48) make decisions. In addition, the TEA-SIM tool isa decision support tool that can be adopted by variousindustries to model various entities such as customers,products and services. It also provides a medium forcompanies to see the interaction process between agents andhow the interactions influence agent decisions. The resultderived from this process can be used to provide informationto organisations so that they can strengthen their customerrelationship management (CRM) strategies by exploring theeffect of word of mouth to improve customer retention.

The dynamic nature of the TEA-SIM tool provides aunique approach to understanding customer behaviour. Itcan be used to observe the pattern of customer interactionand perform further analytics to explore the possibilities ofincorporating those patterns into the marketing strategies inorder to increase revenues. The TEA-SIM tool also worksas a generic model that captures the key drivers behindcustomer change of behaviour and it can also work well ina consultancy environment. It is a cross-industry tool thatis not unique to any industry. Experimenting with the TEA-SIM tool proves that it can improve understanding on factorsthat can drive customer churn. The TEA-SIM tool is not aprecise prediction tool. As a result, agent-based modelling isused to provide insight into the behaviour of a population ofcustomers.

Applying the CADET Approach to a Real-WorldDatasetModels are built for the purpose of mimicking real-worldevents. The CADET approach for conducting agent basedmodelling is a novel approach for building agent basedmodels using decision trees. In this study, the decision treeanalysis is used to describe customers that either remain withtheir mobile network operator (MNO) after their contractperiod and customers that leave their MNO after theircontract period. The CADET approach shows that customersdecision to stay or leave their MNO is driven by a set ofsubjective attributes.The attributes are the individual nodeson the decision tree. Mobile network customers may haveattributes such as mobile phone type, customer location,



Name DescriptionContract length Length of contractGender Customer’s genderSales channel Company that delivered contractPost code Post code in which customer livesCounty Name The name of county where customer livesRegion The region where customer livesDevices Name and model of device used by

customerTenure Number of months with the present mobile

service providerLife stage segment Customer ageNumber ofcomplaints

Number of complaints throughout contactperiod

Q2 bytes 2nd Quarter Data UsageQ3 bytes Third Quarter Data UsageQ2 Voice 2nd Quarter Voice UsageQ3 Voice 3rd Quarter Voice UsageNo of Repairs Number of times customer phone has been

repairedProb Handset Number of Times Customer has reported

problems with handsetTable 2. Dataset description

price plan and data usage. The decision to churn or stay iscomposed of a number of phases represented on the decisiontree. The final node of the decision tree is customer finaldecision to stay with or leave a MNO. Decision trees consistof dependent and independent variables. The dependentvariables are the nodes that make up the final node. Thesevariables influence customers final decision. The primarypurpose of applying the CADET approach for ABMSusing the TEA-SIM application is to understand how muchinfluence a customer’s environment, family, and friendswithin their network have on customer retention. In addition,the CADET approach utilised with the TEA-SIM modelprovides information on the possible decisions a customermight make when they interact with other customers withintheir network or environment. We believe that if a customermeets another customer within the environment, and theyshare the same MNO, they may have a conversation ontheir overall customer experience and that conversation mayinfluence a customers decision to renew their contract. TheCADET approach to agent based modelling seeks to providean understanding on how customer variables along withcustomer network can influence customer retention. The nextsection presents a description of the dataset used to conductthe experiment in this study.

Dataset DescriptionTo demonstrate the usability of the CADET approach, wecollected a dataset from a UK telecommunications company.The dataset comprises of 19,919 observations and 16variables. The dependent variable (output variable) for thisdataset is whether the customer churned or stayed with theirmobile Network operator after their contract. The predictorvariables (input variables) are customers’ data (such as typeof device, price plan and region). The dataset contains 50%of customers who churned and 50% of customers who stayed

Step 1 If tenure is > or < 24 consider the nextnode.

Step 2 If tenure is ≯ 24 and tenure < 14, churn,else consider the next node.

Step 3 If tenure ≯ 24, and tenure ≮ 24, and device= samsung then churn, else renew contract.

Step 4 If tenure > 24, and contract length ≯ 12,then churn else consider the next node.

Step 5 If tenure > 24, and contract length > 12,and region = midlands, then churn, elseconsider the next node.

Step 6 If tenure > 24, and contract length >12, and region 6= midlands, and region 6=London, then churn else consider the nextnode.

Step 7 If tenure > 24, and contract length >12, and region 6= midlands, and region =London, and device = iphone 5, then churn,else move to the next node.

Step 8 If tenure > 24, and contract length >12, and region 6= midlands, and region= London, and device = iphone 5, andproblem with handset > 3, and gender =male, then churn else renew contract.

Step 9 If tenure > 24, and contract length >12, and region 6= midlands, and region= London, and device = iphone 5, andproblem with handset ≯ 3, and numberof complaints >2, then churn, else renewcontract.

Table 3. Steps for Decision Tree Analysis

with the MNO until the end of their contract. The datasetis based on a 24 month contract. Some customers stayed


Mgbemena and Bell 9

Figure 5. Decision tree analysis

with the MNO after the end of the 24 month period (i.ethey renewed their contract) while other customers left theMNO after the 24 month contract. The dataset comprisesof different data types as a means to represent the entirecustomer base. Table 2 presents a tabular description ofthe dataset. The next section presents details of experimentconducted.

Model Structure & ExperimentTo apply the CADET approach with the TEA-SIM tool, theCART decision tree algorithm is run on the dataset describedabove and the result of the decision tree is visualised. Fromthe decision tree (see figure 5), we can see the flow of thedecision process for customers who have decided to staywith their MNO or move to a different MNO. Leaves of thedecision tree instructs the required agent types (e.g. Churn orRenew Contract) and the earlier branching directs specificagent behaviour. The top of decision tree analysis showsthat a customer’s tenure is either greater than 24 or not.If tenure is greater than 24 then move to the next variableon the left. However, if tenure is not greater than 24, thenmove to the next variable on the right-side of the tree. Thisprocess continues until the end of the tree. The end of thetree displays the end result of the process i.e. either customerchurned or renewed their contract at the end of the 24-monthcontract period. Figure 5 displays the decision tree analysisdiagram.

To simulate this process, agents are fed into the Tea-Sim tool as json data definition. Agents behaviour withinthe ABMS environment is illustrated in figure 6. Agentstypically communicate with other agents in the simulationenvironment and take actions after evaluating services

provided by their MNO. To conduct the ABMS experiment,three files are created and run on the Tea-sim tool. Thefiles are init.json, model.json and steps.php. The json filesdescribe agent attributes and the ABMS environment. Thephp file describes the interaction process of the agents.The init.json file (figure 7) is composed of the followingattributes: grid, agent and simulation. n and m under ”grid”represent the number of rows and columns required for theABMS experiment. ”Agents” is composed of agent type,instances and the position of the agents. The init.json filesshow that the instances are six in number and the positionis random. ”Simulation” on the init.json file represent thenumber of runs for the experiment. In this experiment, ”start”is 0 and ”end” is 25. This means that the agents in the gridshould move 25 times.The model.json file (figure 8) describes customer type andattributes. There are two types of customers with customerid 1 and customer id 2. Customer id 1 represent churncustomers while customer id 2 represent stay customers.From the decision tree analysis carried out (figure 6.4), churncustomers have the following attributes:

tenure >14, device = samsung, contract >12 months,region = midlands, region 6= London, device = iPhone5,problem with handset >3, gender = male, number ofcomplaints >2.

Furthermore, stay customers have the followingattributes: tenure >24, contract length > 12, region6= Midlands, region = London, device6= iPhone5,problem with handset < 3, gender = female, number ofcomplaints < 2.

The steps.php file (figure 9) describes agent interaction.To summarise the code, if a churn customer meets a stay



Figure 6. Conceptual Architecture of Customer Behaviour Modelling

Figure 7. Init.json file

Figure 8. Model.json

customer, the stay morphs to a churn customer. The cryingface in figure 10 represents churn customer and the smileyface represent stay customers. Overall, this experimentshows how customer behaviour can be influenced by theenvironment. Figure 10 shows the process in which theagents interact on the TEA-SIM model. The agents movefrom one grid to another and their decisions are based on theinteraction with family/friends as represented in figure 10.

ConclusionCustomer retention is extremely important to mobile networkoperators because of the fierce competition in the MSI.

Hence, companies in the MSI are increasingly strengtheningthe CRM strategies in order to retain customers. Numerousfactors can influence the decision for customers to purchaseor adopt a product or service. However, customers are likelyto trust the word of mouth of someone in their network.This paper presents the CADET approach, a novel data-driven approach to ABMS, thereby developing artefacts inthe process of study. March and Smith (70) describe artefactsas constructs, models, methods and instantiations. This paperpresents a set or artefacts where each artefact contributesto the process of modelling and simulation. This studypresents the CADET approach as a data-driven frameworkfor ABMS. The CADET approach is utilised with a novelABMS tool, the TEA-SIM tool. The application of theCADET approach on the TEA-SIM tool demonstrates theeffectiveness of the data-driven process used to conduct theABMS experiment. This effectiveness of this experimentdemonstrates instantiation in the form, of a working artefact(70).

The TEA-SIM tool provides a generic approach forcompanies across industries to better understand theircustomers. In addition, the TEA-SIM tool can also helpdecision makers in organisations work on better strategies tounderstand customer behaviour, enhance customer retention,and subsequently improve CRM.

This paper also presents a real-life example of imple-menting the CADET approach using the TEA-SIM tool. TheCADET approach was used to effectively provide an Agent-Based model using the results derived from decision trees.In this study, the CADET approach was applied to a datasetfrom a telecommunications company. However, the conceptof the CADET approach can be extended and implementedin other industries such as health-care, manufacturing andfinancial services.

References

[1] V. Grimm, U. Berger, F. Bastiansen, S. Eliassen, V. Ginot,J. Giske, J. Goss-Custard, T. Grand, S. K. Heinz, G. Huse


Mgbemena and Bell 11

Figure 9. Stepper function for dataset

Figure 10. Agent interaction process

et al., “A standard protocol for describing individual-basedand agent-based models,” Ecological modelling, vol. 198,no. 1, pp. 115–126, 2006.

[2] A. Ozcan, “Mobile phones democratize and cultivate next-generation imaging, diagnostics and measurement tools,” Labon a chip, vol. 14, no. 17, pp. 3187–3194, 2014.

[3] W. Verbeke, K. Dejaeger, D. Martens, J. Hur, and B. Baesens,“New insights into churn prediction in the telecommunicationsector: A profit driven data mining approach,” EuropeanJournal of Operational Research, vol. 218, no. 1, pp. 211–229, 2012.

[4] A. Keramati and S. M. Ardabili, “Churn analysis for an iranianmobile operator,” Telecommunications Policy, vol. 35, no. 4,pp. 344–356, 2011.

[5] A. M. Hughes, Strategic database marketing. McGraw-HillNew York, 2006.

[6] M. A. Jones, V. A. Taylor, and K. E. Reynolds, “The effectof requests for positive evaluations on customer satisfactionratings,” Psychology & Marketing, vol. 31, no. 3, pp. 161–170, 2014.

[7] J. Ganesh, M. J. Arnold, and K. E. Reynolds, “Understandingthe customer base of service providers: an examination ofthe differences between switchers and stayers,” Journal ofmarketing, vol. 64, no. 3, pp. 65–87, 2000.

[8] J. G. Blodgett, K. L. Wakefield, and J. H. Barnes, “The effectsof customer service on consumer complaining behavior,”Journal of Services Marketing, vol. 9, no. 4, pp. 31–42, 1995.

[9] J. Itzkowitz, “Customers and cash: How relationships affectsuppliers’ cash holdings,” Journal of Corporate Finance,vol. 19, pp. 159–180, 2013.

[10] R. T. Rust and A. J. Zahorik, “Customer satisfaction, customerretention, and market share,” Journal of retailing, vol. 69,no. 2, pp. 193–215, 1993.

[11] M. J. Epstein and R. A. Westbrook, “Linking actions to profitsin strategic decision making,” Image, 2012.

[12] C. Baumann, G. Elliott, and S. Burton, “Modeling customersatisfaction and loyalty: survey data versus data mining,”Journal of services marketing, vol. 26, no. 3, pp. 148–157,2012.

[13] M. D. Johnson, A. Herrmann, F. Huber, and A. Gustafsson,Customer retention in the automotive industry: quality,satisfaction and loyalty. Springer Science & Business Media,2012.

[14] G. Fullerton, “The moderating effect of normative commit-ment on the service quality-customer retention relationship,”European Journal of Marketing, vol. 48, no. 3/4, pp. 657–673,2014.

[15] D. Ribeiro Soriano, D. Jyh-Fu Jeng, and T. Bailey, “Assessingcustomer retention strategies in mobile telecommunications:Hybrid mcdm approach,” Management Decision, vol. 50,no. 9, pp. 1570–1595, 2012.

[16] A. A. Khan, S. Jamwal, and M. Sepehri, “Applying datamining to customer churn prediction in an internet serviceprovider,” International Journal of Computer Applications,vol. 9, no. 7, pp. 8–14, 2010.

[17] M. Hanif, S. Hafeez, and A. Riaz, “Factors affecting customersatisfaction,” International Research Journal of Finance andEconomics, vol. 60, pp. 44–52, 2010.

[18] K. Silva and S. Yapa, “Customer retention: with specialreference to telecommunication industry in sri lanka,” 2013.

[19] G. B. Svendsen and N. K. Prebensen, “The effect of brand onchurn in the telecommunications sector,” European Journal ofMarketing, vol. 47, no. 8, pp. 1177–1189, 2013.

[20] B.-H. Chu, M.-S. Tsai, and C.-S. Ho, “Toward a hybriddata mining model for customer retention,” Knowledge-BasedSystems, vol. 20, no. 8, pp. 703–718, 2007.



[21] C. Phadke, H. Uzunalioglu, V. B. Mendiratta, D. Kushnir,and D. Doran, “Prediction of subscriber churn using socialnetwork analysis,” Bell Labs Technical Journal, vol. 17, no. 4,pp. 63–75, 2013.

[22] W. Verbeke, D. Martens, and B. Baesens, “Social networkanalysis for customer churn prediction,” Applied SoftComputing, vol. 14, pp. 431–446, 2014.

[23] M.-O. Richard and J.-C. Chebat, “Modeling online consumerbehavior: Preeminence of emotions and moderating influencesof need for cognition and optimal stimulation level,” Journalof Business Research, vol. 69, no. 2, pp. 541–553, 2016.

[24] A. Vag, “Simulating changing consumer preferences: adynamic conjoint model,” Journal of Business Research,vol. 60, no. 8, pp. 904–911, 2007.

[25] C. Rygielski, J.-C. Wang, and D. C. Yen, “Data mining tech-niques for customer relationship management,” Technology insociety, vol. 24, no. 4, pp. 483–502, 2002.

[26] S. Ottar Olsen, A. Alina Tudoran, K. Brunsø, and W. Verbeke,“Extending the prevalent consumer loyalty modelling: the roleof habit strength,” European Journal of Marketing, vol. 47, no.1/2, pp. 303–323, 2013.

[27] M. Xu and J. Walton, “Gaining customer knowledge throughanalytical crm,” Industrial management & data systems, vol.105, no. 7, pp. 955–971, 2005.

[28] S. A. Neslin, S. Gupta, W. Kamakura, J. Lu, and C. H.Mason, “Defection detection: Measuring and understandingthe predictive accuracy of customer churn models,” Journalof marketing research, vol. 43, no. 2, pp. 204–211, 2006.

[29] P. Furness, “Techniques for customer modelling in crm,”Journal of Financial Services Marketing, vol. 5, no. 4, pp.293–307, 2001.

[30] Y. Huang and T. Kechadi, “An effective hybrid learningsystem for telecommunication churn prediction,” ExpertSystems with Applications, vol. 40, no. 14, pp. 5635–5647,2013.

[31] A. T. Jahromi, S. Stakhovych, and M. Ewing, “Customerchurn models: A comparison of probability and data miningapproaches,” in Looking Forward, Looking Back: Drawing onthe Past to Shape the Future of Marketing. Springer, 2016,pp. 144–148.

[32] Z. Zhang, R. Wang, W. Zheng, S. Lan, D. Liang, andH. Jin, “Profit maximization analysis based on data miningand the exponential retention model assumption with respectto customer churn problems,” in 2015 IEEE InternationalConference on Data Mining Workshop (ICDMW). IEEE,2015, pp. 1093–1097.

[33] Y. Ye, S. Liu, and J. Li, “A multiclass machine learningapproach to credit rating prediction,” in InformationProcessing (ISIP), 2008 International Symposiums on. IEEE,2008, pp. 57–61.

[34] A. Lemmens and C. Croux, “Bagging and boostingclassification trees to predict churn,” Journal of MarketingResearch, vol. 43, no. 2, pp. 276–286, 2006.

[35] B. Huang, M. T. Kechadi, and B. Buckley, “Customerchurn prediction in telecommunications,” Expert Systems withApplications, vol. 39, no. 1, pp. 1414–1425, 2012.

[36] J. Burez and D. Van den Poel, “Separating financial from com-mercial customer churn: A modeling step towards resolvingthe conflict between the sales and credit department,” ExpertSystems with Applications, vol. 35, no. 1, pp. 497–514, 2008.

[37] C. Lee, N. Kwak, and C. Lee, “Understanding consumerchurning behaviors in mobile telecommunication serviceindustry: Cross-national comparison between korea andchina,” 2015.

[38] B. He, Y. Shi, Q. Wan, and X. Zhao, “Prediction of customerattrition of commercial banks based on svm model,” ProcediaComputer Science, vol. 31, pp. 423–430, 2014.

[39] P. S. Raju, V. R. Bai, and G. K. Chaitanya, “Data mining:Techniques for enhancing customer relationship managementin banking and retail industries,” International Journalof Innovative Research in Computer and CommunicationEngineering, vol. 2, no. 1, pp. 2650–2657, 2014.

[40] G. G. Sundarkumar and V. Ravi, “A novel hybridundersampling method for mining unbalanced datasets inbanking and insurance,” Engineering Applications of ArtificialIntelligence, vol. 37, pp. 368–377, 2015.

[41] V. L. Migueis, D. Van den Poel, A. S. Camanho, and J. F.e Cunha, “Modeling partial customer churn: On the valueof first product-category purchase sequences,” Expert systemswith applications, vol. 39, no. 12, pp. 11 250–11 256, 2012.

[42] V. L. Migueis, A. Camanho, and J. F. e Cunha, “Customerattrition in retailing: An application of multivariate adaptiveregression splines,” Expert Systems with Applications, vol. 40,no. 16, pp. 6225–6232, 2013.

[43] Y. Richter, E. Yom-Tov, and N. Slonim, “Predicting customerchurn in mobile networks through analysis of social groups.”in SDM, vol. 2010. SIAM, 2010, pp. 732–741.

[44] J. Scott, Social network analysis. Sage, 2012.[45] M. Hassouna, A. Tarhini, T. Elyas, and M. S. AbouTrab,

“Customer churn in mobile markets: a comparison oftechniques,” International Business Research, vol. 8, no. 6,p. 224, 2015.

[46] D. J. Watts, “The” new” science of networks,” Annual reviewof sociology, pp. 243–270, 2004.

[47] C. M. Macal and M. J. North, “Tutorial on agent-basedmodelling and simulation,” Journal of simulation, vol. 4,no. 3, pp. 151–162, 2010.

[48] S. Milgram, “The small world problem,” Psychology today,vol. 2, no. 1, pp. 60–67, 1967.

[49] N. Baxter, D. Collings, and I. Adjali, “Agent-basedmodellingintelligent customer relationship management,” BTTechnology Journal, vol. 21, no. 2, pp. 126–132, 2003.

[50] X. Zhang, Z. Liu, X. Yang, W. Shi, and Q. Wang, “Predictingcustomer churn by integrating the effect of the customercontact network,” in Service Operations and Logistics andInformatics (SOLI), 2010 IEEE International Conference on.IEEE, 2010, pp. 392–397.

[51] H. Risselada, P. C. Verhoef, and T. H. Bijmolt, “Dynamiceffects of social influence and direct marketing on theadoption of high-technology products,” Journal of Marketing,vol. 78, no. 2, pp. 52–68, 2014.

[52] Q. Luo and D. Zhong, “Using social network analysisto explain communication characteristics of travel-relatedelectronic word-of-mouth on social networking sites,”Tourism Management, vol. 46, pp. 274–282, 2015.

[53] E. W. Anderson, “Customer satisfaction and word of mouth,”Journal of service research, vol. 1, no. 1, pp. 5–17, 1998.

[54] S. M. Keaveney, “Customer switching behavior in serviceindustries: An exploratory study,” The Journal of Marketing,


Mgbemena and Bell 13

pp. 71–82, 1995.[55] M. B. Hassouna, “Agent based modelling and simulation: An

examination of customer retention in the uk mobile market,”Ph.D. dissertation, Brunel University, School of InformationSystems, Computing and Mathematics, 2012.

[56] P. Twomey and R. Cadman, “Agent-based modelling ofcustomer behaviour in the telecoms and media markets,” info,vol. 4, no. 1, pp. 56–63, 2002.

[57] M. J. North, C. M. Macal, J. S. Aubin, P. Thimmapuram,M. Bragen, J. Hahn, J. Karr, N. Brigham, M. E. Lacy,and D. Hampton, “Multiscale agent-based consumer marketmodeling,” Complexity, vol. 15, no. 5, pp. 37–47, 2010.

[58] J. D. Farmer and D. Foley, “The economy needs agent-basedmodelling,” Nature, vol. 460, no. 7256, pp. 685–686, 2009.

[59] J. M. Epstein, “Modelling to contain pandemics,” Nature, vol.460, no. 7256, pp. 687–687, 2009.

[60] A. J. Heppenstall, A. T. Crooks, L. M. See, and M. Batty,Agent-based models of geographical systems. SpringerScience & Business Media, 2011.

[61] C. Adami, Introduction to artificial life. Springer Science &Business Media, 1998, vol. 1.

[62] T. Zhang and D. Zhang, “Agent-based simulation of consumerpurchase decision-making and the decoy effect,” Journal ofBusiness Research, vol. 60, no. 8, pp. 912–922, 2007.

[63] G. N. Gilbert, Agent-based models. Sage, 2008, no. 153.[64] O. Roozmand, N. Ghasem-Aghaee, G. J. Hofstede, M. A.

Nematbakhsh, A. Baraani, and T. Verwaart, “Agent-basedmodeling of consumer decision making process based onpower distance and personality,” Knowledge-Based Systems,vol. 24, no. 7, pp. 1075–1095, 2011.

[65] L. An, “Modeling human decisions in coupled human andnatural systems: review of agent-based models,” EcologicalModelling, vol. 229, pp. 25–36, 2012.

[66] H. R. Gimblett, Integrating geographic information systemsand agent-based modeling techniques for simulating socialand ecological processes. Oxford University Press NewYork, 2002, vol. 168.

[67] P. Deadman, D. Robinson, E. Moran, and E. Brondizio,“Colonist household decisionmaking and land-use changein the amazon rainforest: an agent-based simulation,”Environment and Planning B, vol. 31, pp. 693–710, 2004.

[68] V. K. Vaishnavi and W. Kuechler, Design science researchmethods and patterns: innovating information and communi-cation technology. Crc Press, 2015.

[69] R. H. von Alan, S. T. March, J. Park, and S. Ram, “Designscience in information systems research,” MIS quarterly,vol. 28, no. 1, pp. 75–105, 2004.

[70] S. T. March and G. F. Smith, “Design and natural scienceresearch on information technology,” Decision supportsystems, vol. 15, no. 4, pp. 251–266, 1995.

[71] A. Newell, H. A. Simon et al., Human problem solving.Prentice-Hall Englewood Cliffs, NJ, 1972, vol. 104, no. 9.

[72] K. Peffers, T. Tuunanen, M. A. Rothenberger, and S. Chatter-jee, “A design science research methodology for informationsystems research,” Journal of management information sys-tems, vol. 24, no. 3, pp. 45–77, 2007.

[73] B. Kuechler and V. Vaishnavi, “On theory development indesign science research: anatomy of a research project,”European Journal of Information Systems, vol. 17, no. 5, pp.

489–504, 2008.[74] M. Hassouna and M. Arzoky, “Agent based modelling and

simulation: Toward a new model of customer retention inthe mobile market,” in Proceedings of the 2011 SummerComputer Simulation Conference. Society for Modeling &Simulation International, 2011, pp. 30–35.

[75] J. G. Barnes and D. M. Howlett, “Predictors of equity inrelationships between financial services providers and retailcustomers,” International Journal of Bank Marketing, vol. 16,no. 1, pp. 15–23, 1998.

[76] L. Rokach and O. Maimon, Data mining with decision trees:theory and applications. World scientific, 2014.

[77] S. B. Kotsiantis, “Decision trees: a recent overview,” ArtificialIntelligence Review, vol. 39, no. 4, pp. 261–283, 2013.

[78] C.-F. Tsai and Y.-H. Lu, “Data mining techniques in customerchurn prediction,” Recent Patents on Computer Science,vol. 3, no. 1, pp. 28–32, 2010.

[79] C. A. Iglesias, M. Garijo, and J. C. Gonzalez, “A survey ofagent-oriented methodologies,” in Intelligent Agents V: AgentsTheories, Architectures, and Languages. Springer, 1999, pp.317–330.

[80] D. Bell, A. Kashefi, N. Saleh, and T. Turchi, “A data-driven agent based simulation platform for early healtheconomics device evaluation,” to appear in Proceedings of theSpringSim’16, 2016.


journal title a data-driven approach to agent based ... · simulation (abms) as a novel approach to...

Documents